{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport matplotlib.pylot as plt\nimport torch\nimport os\nimport pytorch_lightning as pl\nfrom torch.utils.data import Dataset, Dataloader\nfrom sklearn import model_selection\nimport torchvision.transforms as transforms\nimport torchvision.io\nimport librosa\nfrom PIL import Image\nimport albumentations as alb\nimport torch.multiprocessing as mp\nimport warnings\n\nwarnings.filterwarnings('ignore')","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from pytorch_lightning.callbacks import ModelCheckpoint, BackboneFinetuning, EarlyStopping","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pip install -q torchtoolbox timm","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class Config:\n    use_aug = False\n    num_classes = 264\n    batch_size = 64\n    epochs = 12\n    PRECISION = 16\n    PATIENCE= 8\n    seed = 2023\n    model = 'tf-efficientnet_b0_ns'\n    pretrained =True\n    weight_decay = 1e-3\n    use_mixup = True\n    mixup_aplha = 0.2\n    DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n    \n    data_root = '/kaggle/input/birdclef-2023/'\n    train_path = '/kaggle/input/bc2023-train-val-df/train.csv'\n    valid_path = '/kaggle/input/bc2023-train-val-df/valid.csv'\n    train_images = '/kaggle/input/split-creating-melspecs-stage-1/specs/train/'\n    valid_images = '/kaggle/input/split-creating-melspecs-stage-1/specs/valid/'\n    \n    SR = 32000\n    DURATION = 5\n    MAX_READ_SAMPLES = 5\n    LR = 5e-4","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"pl.seed_everthing(Config.seed, workers= True)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def config_to_dict(cfg):\n    return dict((name, getattr(cfg, name)) for name in dir(cfg) if not name.startswith('__'))","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df_train = pd.read_csv(Config.train_path)\ndf_valid = pd.read_csv(Config.valid_path)\ndf_train.head()","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"Config.num_classes = len(df_train.primary_label.unique())","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df_train = pd.concat([df_train, pd.get_dummies(df_train['primary_label'])], axis=1)\ndf_valid = pd.concat([df_train, pd.get_dummies(df_valid['primary_label'])], axis=1)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"birds = list(df_train.primary_label.unique())","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"missing_birds = list(set(list(df_train.primary_label.unique())).difference(list(df_valid.primary_label.unique())))","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"non_missing_birds = list(set(list(df_train.primary_label.unique())).difference(missing_birds))","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"len(non_missing_birds)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df_valid[missing_birds] = 0\ndf_valid = df_valid[df_train.columns] ## Fix order","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import albumentations as A\ndef get_train_transform():\n    return A.Compose([\n        A.HorizontalFlip(p=0.5),\n        A.OneOf([\n                A.Cutout(max_h_size=5, max_w_size=16),\n                A.CoarseDropout(max_holes=4),\n            ], p=0.5),\n    ])","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class BirdDataset(torch.utils.data.Dataset):\n    \n    def __init__(self, df, sr= Config.SR, duration = Config.DURATION, augmentations= None, train =True):\n        self.df= df\n        self.sr =sr\n        self.train = train\n        self.duration = duration\n        self.augmentations = augmentations\n        if train:\n            self.img_dir = Config.train_images\n        else:\n            self.img_dir = Config.valid_images\n            \n    def __len__(self):\n        return len(self, df)\n    \n    @staticmethod\n    def nirmalize(image):\n        image = image / 255.0\n        return image\n    \n    def __getitem__(self, idx):\n        row = self.df.iloc[idx]\n        impath - self.img_dir + f\"{row.filename}.npy\"\n        \n        image = np.load(str(impath))[:Config.MAX_READ_SAMPLES]\n        \n        if self.train:\n            image = image[np.random.choice(len(image))]\n        else:\n            image == image[0]\n            \n        image = torch.tensor(image).float()\n        if self.augmentations:\n            image = self.augmentations(image.unsqueeze(0)).squeeze()\n        \n        image.size()\n        image = torch.stack([image,image,image])\n        image = self.normalize(image)\n        return image, torch.tensor(row[17:]).float()","metadata":{"execution":{"iopub.status.busy":"2023-06-04T10:49:05.495716Z","iopub.execute_input":"2023-06-04T10:49:05.496520Z","iopub.status.idle":"2023-06-04T10:49:05.530462Z","shell.execute_reply.started":"2023-06-04T10:49:05.496476Z","shell.execute_reply":"2023-06-04T10:49:05.528853Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def get_fold_dls(df_train, df_valid):\n\n    ds_train = BirdDataset(\n        df_train, \n        sr = Config.SR,\n        duration = Config.DURATION,\n        augmentations = None,\n        train = True\n    )\n    ds_val = BirdDataset(\n        df_valid, \n        sr = Config.SR,\n        duration = Config.DURATION,\n        augmentations = None,\n        train = False\n    )\n    dl_train = DataLoader(ds_train, batch_size=Config.batch_size , shuffle=True, num_workers = 2)    \n    dl_val = DataLoader(ds_val, batch_size=Config.batch_size, num_workers = 2)\n    return dl_train, dl_val, ds_train, ds_val","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def show_batch(img_ds, num_items, num_rows, num_cols, predict_arr=None):\n    fig = plt.figure(figsize=(12, 6))    \n    img_index = np.random.randint(0, len(img_ds)-1, num_items)\n    for index, img_index in enumerate(img_index):  # list first 9 images\n        img, lb = img_ds[img_index]        \n        ax = fig.add_subplot(num_rows, num_cols, index + 1, xticks=[], yticks=[])\n        if isinstance(img, torch.Tensor):\n            img = img.detach().numpy()\n        if isinstance(img, np.ndarray):\n            img = img.transpose(1, 2, 0)\n            ax.imshow(img)        \n            \n        title = f\"Spec\"\n        ax.set_title(title) ","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"dl_train, dl_val, ds_train, ds_val = get_fold_dls(df_train, df_valid)\nshow_batch(ds_val, 8, 2, 4)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from torch.optim.lr_scheduler import CosineAnnealingLR, CosineAnnealingWarmRestarts, ReduceLROnPlateau, OneCycleLR\n\ndef get_optimizer(lr, params):\n    model_optimizer = torch.optim.Adam(\n            filter(lambda p: p.requires_grad, params), \n            lr=lr,\n            weight_decay=Config.weight_decay\n        )\n    interval = \"epoch\"\n    \n    lr_scheduler = CosineAnnealingWarmRestarts(\n                            model_optimizer, \n                            T_0=Config.epochs, \n                            T_mult=1, \n                            eta_min=1e-6, \n                            last_epoch=-1\n                        )\n\n    return {\n        \"optimizer\": model_optimizer, \n        \"lr_scheduler\": {\n            \"scheduler\": lr_scheduler,\n            \"interval\": interval,\n            \"monitor\": \"val_loss\",\n            \"frequency\": 1\n        }\n    }","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from torchtoolbox.tools import mixup_data, mixup_criterion\nimport torch.nn as nn\nfrom torch.nn.functional import cross_entropy\nimport torchmetrics\nimport timm","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import sklearn.metrics\n\ndef padded_cmap(solution, submission, padding_factor=5):\n    solution = solution#.drop(['row_id'], axis=1, errors='ignore')\n    submission = submission#.drop(['row_id'], axis=1, errors='ignore')\n    new_rows = []\n    for i in range(padding_factor):\n        new_rows.append([1 for i in range(len(solution.columns))])\n    new_rows = pd.DataFrame(new_rows)\n    new_rows.columns = solution.columns\n    padded_solution = pd.concat([solution, new_rows]).reset_index(drop=True).copy()\n    padded_submission = pd.concat([submission, new_rows]).reset_index(drop=True).copy()\n    score = sklearn.metrics.average_precision_score(\n        padded_solution.values,\n        padded_submission.values,\n        average='macro',\n    )\n    return score\n\ndef map_score(solution, submission):\n    solution = solution#.drop(['row_id'], axis=1, errors='ignore')\n    submission = submission#.drop(['row_id'], axis=1, errors='ignore')\n    score = sklearn.metrics.average_precision_score(\n        solution.values,\n        submission.values,\n        average='micro',\n    )\n    return score","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"dummy = df_valid[birds].copy()\ndummy[birds] = np.random.rand(dummy.shape[0],dummy.shape[1])","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"padded_cmap(df_valid[birds], dummy[birds], padding_factor = 5)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"padded_cmap(df_valid[birds], dummy[birds], padding_factor = 1)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"map_score(df_valid[birds], dummy[birds])","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class BirdClefModel(pl.LightningModule):\n    def __init__(self, model_name=Config.model, num_classes = Config.num_classes, pretrained = Config.pretrained):\n        super().__init__()\n        self.num_classes = num_classes\n\n        self.backbone = timm.create_model(model_name, pretrained=pretrained)\n\n        if 'res' in model_name:\n            self.in_features = self.backbone.fc.in_features\n            self.backbone.fc = nn.Linear(self.in_features, num_classes)\n        elif 'dense' in model_name:\n            self.in_features = self.backbone.classifier.in_features\n            self.backbone.classifier = nn.Linear(self.in_features, num_classes)\n        elif 'efficientnet' in model_name:\n            self.in_features = self.backbone.classifier.in_features\n            self.backbone.classifier = nn.Sequential(\n                nn.Linear(self.in_features, num_classes)\n            )\n        \n        self.loss_function = nn.BCEWithLogitsLoss() \n\n    def forward(self,images):\n        logits = self.backbone(images)\n        return logits\n        \n    def configure_optimizers(self):\n        return get_optimizer(lr=Config.LR, params=self.parameters())\n\n    def train_with_mixup(self, X, y):\n        X, y_a, y_b, lam = mixup_data(X, y, alpha=Config.mixup_alpha)\n        y_pred = self(X)\n        loss_mixup = mixup_criterion(cross_entropy, y_pred, y_a, y_b, lam)\n        return loss_mixup\n\n    def training_step(self, batch, batch_idx):\n        image, target = batch        \n        if Config.use_mixup:\n            loss = self.train_with_mixup(image, target)\n        else:\n            y_pred = self(image)\n            loss = self.loss_function(y_pred,target)\n\n        self.log(\"train_loss\", loss, on_step=True, on_epoch=True, prog_bar=True)\n        return loss        \n\n    def validation_step(self, batch, batch_idx):\n        image, target = batch     \n        y_pred = self(image)\n        val_loss = self.loss_function(y_pred, target)\n        self.log(\"val_loss\", val_loss, on_step=True, on_epoch=True, logger=True, prog_bar=True)\n        \n        return {\"val_loss\": val_loss, \"logits\": y_pred, \"targets\": target}\n    \n    def train_dataloader(self):\n        return self._train_dataloader \n    \n    def validation_dataloader(self):\n        return self._validation_dataloader\n    \n    def validation_epoch_end(self,outputs):\n        avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()\n        output_val = torch.cat([x['logits'] for x in outputs],dim=0).sigmoid().cpu().detach().numpy()\n        target_val = torch.cat([x['targets'] for x in outputs],dim=0).cpu().detach().numpy()\n        \n        # print(output_val.shape)\n        val_df = pd.DataFrame(target_val, columns = birds)\n        pred_df = pd.DataFrame(output_val, columns = birds)\n        \n        avg_score = padded_cmap(val_df, pred_df, padding_factor = 5)\n        avg_score2 = padded_cmap(val_df, pred_df, padding_factor = 3)\n        avg_score3 = sklearn.metrics.label_ranking_average_precision_score(target_val,output_val)\n        \n#         competition_metrics(output_val,target_val)\n        print(f'epoch {self.current_epoch} validation loss {avg_loss}')\n        print(f'epoch {self.current_epoch} validation C-MAP score pad 5 {avg_score}')\n        print(f'epoch {self.current_epoch} validation C-MAP score pad 3 {avg_score2}')\n        print(f'epoch {self.current_epoch} validation AP score {avg_score3}')\n        \n        \n        val_df.to_pickle('val_df.pkl')\n        pred_df.to_pickle('pred_df.pkl')\n        \n        \n        return {'val_loss': avg_loss,'val_cmap':avg_score}","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from pytorch_lightning.loggers import WandbLogger\nimport gc\n\ndef run_training():\n    print(f\"Running training...\")\n    logger = None\n    \n    \n    dl_train, dl_val, ds_train, ds_val = get_fold_dls(df_train, df_valid)\n    \n    audio_model = BirdClefModel()\n\n    early_stop_callback = EarlyStopping(monitor=\"val_loss\", min_delta=0.00, patience=Config.PATIENCE, verbose= True, mode=\"min\")\n    checkpoint_callback = ModelCheckpoint(monitor='val_loss',\n                                          dirpath= \"/kaggle/working/exp1/\",\n                                      save_top_k=1,\n                                      save_last= True,\n                                      save_weights_only=True,\n                                      filename= f'./{Config.model}_loss',\n                                      verbose= True,\n                                      mode='min')\n    \n    callbacks_to_use = [checkpoint_callback,early_stop_callback]\n\n\n    trainer = pl.Trainer(\n        gpus=1,\n        val_check_interval=0.5,\n        deterministic=True,\n        max_epochs=Config.epochs,\n        logger=logger,\n        auto_lr_find=False,    \n        callbacks=callbacks_to_use,\n        precision=Config.PRECISION, accelerator=\"gpu\" \n    )\n\n    print(\"Running trainer.fit\")\n    trainer.fit(audio_model, train_dataloaders = dl_train, val_dataloaders = dl_val)                \n\n    gc.collect()\n    torch.cuda.empty_cache()","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"run_training()","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"pred = pd.read_pickle('/kaggle/working/pred_df.pkl')\ntrue = pd.read_pickle('/kaggle/working/val_df.pkl')","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"true.sum(axis=1)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"pred.sum(axis=1)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"padded_cmap(true, pred, padding_factor = 5)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}